Abstract
Myocardial echocardiography segmentation is useful for cardiac function assessment, yet remains challenging due to continuous myocardial motion and deformation during the cardiac cycle, as well as boundary ambiguity caused by speckle noise. Existing memory-based segmentation methods typically construct memory banks by accumulating single-frame representations, where the stored memory items fail to explicitly capture motion dynamics across consecutive frames. Meanwhile, spatiotemporal Transformer-based methods often rely on full-sequence, limiting their applicability to online scenarios where frames arrive sequentially. In this work, we propose a Motion memory Spatiotemporal Transformer Network (MoST-Net) for myocardial echocardiography segmentation, MoST-Net integrates motion modulation into long-term memory learning, enabling the long-term memory to be updated as frames arrive while continuously accumulating temporal dependencies. Specifically, motion memory learning leverages progressively estimated motion states to modulate image features and capture short-term motion memory, which are further extended to long-range dependencies through long-term memory modeling. The memory prompt encoder maps both image and memory features into motion prompts, which are adapted by a spatiotemporal decoder to obtain the final segmentation, and the boundary uncertainty enhancement module is introduced within the decoder to alleviate ambiguous myocardial boundaries. In addition, the adaptive ranking strategy maintains an effective memory length by jointly considering temporal consistency and segmentation quality. Experiments on CAMUS, HMC-QU, and a private myocardial dataset demonstrate that MoST-Net outperforms state-of-the-art methods even under severe noise. Further analysis shows that longer memory banks may not be necessary when motion-aware memory learning is employed, demonstrating its potential for online myocardial segmentation in clinical scenarios.